Integrating the Pre-trained Item Representations with Reformed Self-attention Network for Sequential Recommendation
Guanzhong Liang, Jie Liao, Wei Guo Zhou, Junhao Wen · 2022
The conventional sequential recommendation tends to recommend an item to a sequence due to the concurrence of this item in other similar sequences. It is a typical collaborative filtering recommendation strategy that suffers from the data sparsity problem. Integrating the item-side information into the recommendation model can alleviate the data sparsity problem and avoid the privacy issue. However, roughly fusing item-side information in item embeddings to optimize for final recommendation accuracy might lose the objective associations among items. Thus, this paper proposes a self-supervised learning task on item-side information to generate the item representation space in advance of the recommendation task. The space can gather similar items and separate the different items. Then we reform a recently successful sequential recommendation method, self-attention network, by integrating an untrainable mechanism to fully take advantage of the pre-trained item representation space. The two procedures comprise a hybrid recommendation framework to achieve collaborative filtering and content-based recommendation strategies. Extensive experiments conducted on two benchmark datasets demonstrate the superiority of our proposed methods over other baseline methods.